As we develop AI agents, here are a few important considerations we’ve discovered:
Current APIs are designed for humans driven workflows
AI agents need APIs to navigate and access application data, but many existing APIs are built with human-assisted workflows in mind. This limits the agent’s performance potential because the APIs may not expose direct, data-focused endpoints suited for autonomous use.
Screen-based interaction can be slower
AI agents can leverage Computer Use Automation (CUA) by analyzing screenshots or the DOM, but this often makes them slower than humans. The agent must capture and interpret each view before proceeding, whereas a human user very quickly click through familiar screens.
Clear UI design helps agents avoid wandering
If screen layouts and navigation paths are ambiguous or inconsistent, AI agents get stuck or take non-optimal routes. Humans typically learn shortcuts and best practices after a few uses, but an AI agent has to build up its own understanding over time.
Context is everything
Every action an AI agent takes depends on the context you provide—previous steps, screenshots, or chat history. Storing and referencing this context can get computationally heavy, but it’s essential for maintaining continuity and improving performance over repeated tasks.
Security and Auditing
AI agents acting on behalf of users can introduce auditing or compliance complications. Some systems aren’t prepared to log “AI user” actions distinctly from those of real human users.
Memory & Learning
Over time, an AI agent can build better strategies if it retains logs of previous actions and outcomes. This can eventually yield very efficient navigation patterns—almost as good or better than seasoned human users.
AI agents challenge existing paradigms
While AI agents hold tremendous promise, they also raise questions about how we design APIs, create user experiences, and handle role-based access control (RBAC). We may need to rethink these foundations for AI-driven workflows to truly excel.
What do you think?
